Classification Of Land Categories Based On High-resolution Remote Sensing Images

This paper used Longyuan street of Longquan, Zhejiang as the study area and SPOT5 remote sensing images of study area as the data source, conducted a research of classification experiment.This paper conducted the classification research by different methods to the image in 2009. Then compared and analyzed the classification results based on the data of forest resources management survey in 2009. With the classification results,We chose the Object-oriented classification method to classify the image in 2011, then extracted and analyzed land usage changeful information of two classification results. Obtained the following conclusions:(1) The image preprocessing mainly includes Geometric correction, Radiometric Correction, Image cutting, and Image fusion which includes four methods, they are Brovey, CN, PCA and Gram-Schmidt Pan Sharpening. The results were evaluated by six indexes, they are Standard deviation, Entropy, Average gradient, Correlation coefficient, Deviation index, Spectral distortion degree. Choose the pan fusion result by comparing six fusion results to do research later.(2) Mahalanobis distance classification, maximum likelihood classification, neural networks and support vector machine classification have been used to classify the image. And then establish confusion matrix to obtain the results of accuracy assessment, the finally accuracy assessment results are as follows: The overall accuracy of Mahalanobis distance classification result is 80.2%, and the Kappa coefficient is 0.7072;The overall accuracy of Maximum likelihood classification result is 81.2%, and the Kappa coefficient is 0.7230;The overall accuracy of Neural network classification result is 84%, and the Kappa coefficient is 0.7674; The overall accuracy of support vector machine classification result is 85.8%, and the Kappa coefficient is 0.7872.(3)The multi-scale segmentation experiment by setting different parameters was conducted in the eCognition software. After several experiments, the parameter was chosen as follows: the band weight is 1.54:1.4:1:1.47, the shape factor is 0.2, the compactness factor is 0.5.With these parameters above, we can achieve the segmentation results and objects, then the objects were classified with Closest classifier and membership functions classifier, after making full use of spectral,texture, shape features, and selecting the best feature space. The finally overall accuracy of object-oriented classification result is 91.40%, and the Kappa coefficient is 0.8735.(4)Classify two phase images in year 2009 and 2011 with Objected-orient classification method, and then the two phase classification results were applied to the study of dynamic change of land.The statistical results shows that: The forest land area is 8782.79 ha in 2009,and 8761.95 ha in year 2011;The non-forest land area is 1361.24 ha in 2009,and 1374.22 ha in 2010. The forest land area has reduced, the non-forest area has increased.(5)The subcompartment boundary was extracted based on SPOT5 images in 2009.